| 1 |
// Auto-generated file. Do not edit! |
| 2 |
// Template: src/qs8-igemm/MRx8c8-avxvnni.c.in |
| 3 |
// Generator: tools/xngen |
| 4 |
// |
| 5 |
// Copyright 2024 Google LLC |
| 6 |
// |
| 7 |
// This source code is licensed under the BSD-style license found in the |
| 8 |
// LICENSE file in the root directory of this source tree. |
| 9 |
|
| 10 |
#include <assert.h> |
| 11 |
|
| 12 |
#include <immintrin.h> |
| 13 |
|
| 14 |
#include "xnnpack/common.h" |
| 15 |
#include "xnnpack/gemm.h" |
| 16 |
#include "xnnpack/intrinsics-polyfill.h" |
| 17 |
#include "xnnpack/math.h" |
| 18 |
#include "xnnpack/unaligned.h" |
| 19 |
|
| 20 |
|
| 21 |
void xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_3x8c8__avxvnni( |
| 22 |
size_t mr, |
| 23 |
size_t nc, |
| 24 |
size_t kc, |
| 25 |
size_t ks, |
| 26 |
const int8_t** restrict a, |
| 27 |
const void* restrict w, |
| 28 |
int8_t* restrict c, |
| 29 |
size_t cm_stride, |
| 30 |
size_t cn_stride, |
| 31 |
size_t a_offset, |
| 32 |
const int8_t* zero, |
| 33 |
const union xnn_qs8_qc8w_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS |
| 34 |
{ |
| 35 |
assert(mr != 0); |
| 36 |
assert(mr <= 3); |
| 37 |
assert(nc != 0); |
| 38 |
assert(kc != 0); |
| 39 |
assert(kc % sizeof(int8_t) == 0); |
| 40 |
assert(a != NULL); |
| 41 |
assert(w != NULL); |
| 42 |
assert(c != NULL); |
| 43 |
|
| 44 |
kc = round_up_po2(kc, 8 * sizeof(int8_t)); |
| 45 |
int8_t* c0 = c; |
| 46 |
int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride); |
| 47 |
if XNN_UNPREDICTABLE(mr < 2) { |
| 48 |
c1 = c0; |
| 49 |
} |
| 50 |
int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride); |
| 51 |
if XNN_UNPREDICTABLE(mr <= 2) { |
| 52 |
c2 = c1; |
| 53 |
} |
| 54 |
|
| 55 |
const __m256i vsign_mask = _mm256_set1_epi8(0x80); |
| 56 |
XNN_FORCE_REALIZATION(vsign_mask); |
| 57 |
const __m256 voutput_max_less_zero_point = _mm256_set1_ps((int32_t) params->fp32_scalar.output_max - (int32_t) params->fp32_scalar.output_zero_point); |
| 58 |
const __m256i voutput_zero_point = _mm256_set1_epi32(params->fp32_scalar.output_zero_point); |
| 59 |
const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min); |
| 60 |
// XNN_FORCE_REALIZATION(voutput_max_less_zero_point); |
| 61 |
// XNN_FORCE_REALIZATION(voutput_zero_point); |
| 62 |
// XNN_FORCE_REALIZATION(voutput_min); |
| 63 |
do { |
| 64 |
__m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm_load_si128((const __m128i*) w)); |
| 65 |
__m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm_load_si128((const __m128i*) ((const int32_t*) w + 4))); |
| 66 |
__m256i vacc1x0123 = vacc0x0123; |
| 67 |
__m256i vacc1x4567 = vacc0x4567; |
| 68 |
__m256i vacc2x0123 = vacc0x0123; |
| 69 |
__m256i vacc2x4567 = vacc0x4567; |
| 70 |
w = (const int32_t*) w + 8; |
| 71 |
|
| 72 |
size_t p = ks; |
| 73 |
do { |
| 74 |
const int8_t* restrict a0 = a[0]; |
| 75 |
if XNN_UNPREDICTABLE(a0 != zero) { |
| 76 |
a0 = (const int8_t*) ((uintptr_t) a0 + a_offset); |
| 77 |
} |
| 78 |
const int8_t* restrict a1 = a[1]; |
| 79 |
if XNN_UNPREDICTABLE(a1 != zero) { |
| 80 |
a1 = (const int8_t*) ((uintptr_t) a1 + a_offset); |
| 81 |
} |
| 82 |
const int8_t* restrict a2 = a[2]; |
| 83 |
if XNN_UNPREDICTABLE(a2 != zero) { |
| 84 |
a2 = (const int8_t*) ((uintptr_t) a2 + a_offset); |
| 85 |
} |
| 86 |
a += 3; |
| 87 |
|
| 88 |
size_t k = kc; |
| 89 |
while (k >= 16 * sizeof(int8_t)) { |
| 90 |
const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask); |
| 91 |
const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask); |
| 92 |
a0 += 16; |
| 93 |
const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask); |
| 94 |
const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask); |
| 95 |
a1 += 16; |
| 96 |
const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask); |
| 97 |
const __m256i va2x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask); |
| 98 |
a2 += 16; |
| 99 |
|
| 100 |
const __m256i vb01234567x0123 = _mm256_load_si256(w); |
| 101 |
const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32)); |
| 102 |
const __m256i vb01234567x4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 64)); |
| 103 |
const __m256i vb89ABCDEFx4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 96)); |
| 104 |
|
| 105 |
vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x01234567, vb01234567x0123); |
| 106 |
vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123); |
| 107 |
vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x01234567, vb01234567x0123); |
| 108 |
vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123); |
| 109 |
vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x01234567, vb01234567x0123); |
| 110 |
vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123); |
| 111 |
vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x89ABCDEF, vb01234567x4567); |
| 112 |
vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x89ABCDEF, vb89ABCDEFx4567); |
| 113 |
vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x89ABCDEF, vb01234567x4567); |
| 114 |
vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x89ABCDEF, vb89ABCDEFx4567); |
| 115 |
vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x89ABCDEF, vb01234567x4567); |
| 116 |
vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x89ABCDEF, vb89ABCDEFx4567); |
| 117 |
|
| 118 |
w = (const int8_t*) w + 128; |
| 119 |
k -= 16 * sizeof(int8_t); |
| 120 |
} |
| 121 |
|
| 122 |
if (k != 0) { |
| 123 |
const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask); |
| 124 |
a0 += 8; |
| 125 |
const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask); |
| 126 |
a1 += 8; |
| 127 |
const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask); |
| 128 |
a2 += 8; |
| 129 |
|
| 130 |
const __m256i vb01234567x0123 = _mm256_load_si256(w); |
| 131 |
const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32)); |
| 132 |
|
| 133 |
vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x01234567, vb01234567x0123); |
| 134 |
vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123); |
| 135 |
vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x01234567, vb01234567x0123); |
| 136 |
vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123); |
| 137 |
vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x01234567, vb01234567x0123); |
| 138 |
vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123); |
| 139 |
|
| 140 |
w = (const int8_t*) w + 64; |
| 141 |
k -= 8 * sizeof(int8_t); |
| 142 |
} |
| 143 |
|
| 144 |
p -= 3 * sizeof(void*); |
| 145 |
} while (p != 0); |
| 146 |
|
| 147 |
|
| 148 |
// Add adjacent pairs |
| 149 |
const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567); |
| 150 |
__m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0)); |
| 151 |
const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567); |
| 152 |
__m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0)); |
| 153 |
const __m256i vsum2x02134657 = _mm256_hadd_epi32(vacc2x0123, vacc2x4567); |
| 154 |
__m256i vacc2x01234567 = _mm256_permute4x64_epi64(vsum2x02134657, _MM_SHUFFLE(3, 1, 2, 0)); |
| 155 |
|
| 156 |
__m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567); |
| 157 |
__m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567); |
| 158 |
__m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567); |
| 159 |
|
| 160 |
const __m256 vscale01234567 = _mm256_load_ps(w); |
| 161 |
w = (const float*) w + 8; |
| 162 |
vout0x01234567 = _mm256_mul_ps(vout0x01234567, vscale01234567); |
| 163 |
vout1x01234567 = _mm256_mul_ps(vout1x01234567, vscale01234567); |
| 164 |
vout2x01234567 = _mm256_mul_ps(vout2x01234567, vscale01234567); |
| 165 |
|
| 166 |
vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max_less_zero_point); |
| 167 |
vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max_less_zero_point); |
| 168 |
vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max_less_zero_point); |
| 169 |
|
| 170 |
vacc0x01234567 = _mm256_cvtps_epi32(vout0x01234567); |
| 171 |
vacc1x01234567 = _mm256_cvtps_epi32(vout1x01234567); |
| 172 |
vacc2x01234567 = _mm256_cvtps_epi32(vout2x01234567); |
| 173 |
|
| 174 |
vacc0x01234567 = _mm256_add_epi32(vacc0x01234567, voutput_zero_point); |
| 175 |
vacc1x01234567 = _mm256_add_epi32(vacc1x01234567, voutput_zero_point); |
| 176 |
vacc2x01234567 = _mm256_add_epi32(vacc2x01234567, voutput_zero_point); |
| 177 |
|
| 178 |
vacc0x01234567 = _mm256_packs_epi32(vacc0x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc0x01234567, 1))); |
| 179 |
__m128i voutb0x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc0x01234567, vacc0x01234567)); |
| 180 |
vacc1x01234567 = _mm256_packs_epi32(vacc1x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc1x01234567, 1))); |
| 181 |
__m128i voutb1x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc1x01234567, vacc1x01234567)); |
| 182 |
vacc2x01234567 = _mm256_packs_epi32(vacc2x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc2x01234567, 1))); |
| 183 |
__m128i voutb2x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc2x01234567, vacc2x01234567)); |
| 184 |
|
| 185 |
voutb0x01234567 = _mm_max_epi8(voutb0x01234567, voutput_min); |
| 186 |
voutb1x01234567 = _mm_max_epi8(voutb1x01234567, voutput_min); |
| 187 |
voutb2x01234567 = _mm_max_epi8(voutb2x01234567, voutput_min); |
| 188 |
|
| 189 |
if (nc >= 8) { |
| 190 |
_mm_storel_epi64((__m128i*) c2, voutb2x01234567); |
| 191 |
c2 = (int8_t*) ((uintptr_t) c2 + cn_stride); |
| 192 |
_mm_storel_epi64((__m128i*) c1, voutb1x01234567); |
| 193 |
c1 = (int8_t*) ((uintptr_t) c1 + cn_stride); |
| 194 |
_mm_storel_epi64((__m128i*) c0, voutb0x01234567); |
| 195 |
c0 = (int8_t*) ((uintptr_t) c0 + cn_stride); |
| 196 |
a = (const int8_t**restrict) ((uintptr_t) a - ks); |
| 197 |
nc -= 8; |
| 198 |
} else { |
| 199 |
if (nc & 4) { |
| 200 |
_mm_storeu_si32(c2, voutb2x01234567); |
| 201 |
c2 += 4; |
| 202 |
_mm_storeu_si32(c1, voutb1x01234567); |
| 203 |
c1 += 4; |
| 204 |
_mm_storeu_si32(c0, voutb0x01234567); |
| 205 |
c0 += 4; |
| 206 |
voutb2x01234567 = _mm_srli_epi64(voutb2x01234567, 32); |
| 207 |
voutb1x01234567 = _mm_srli_epi64(voutb1x01234567, 32); |
| 208 |
voutb0x01234567 = _mm_srli_epi64(voutb0x01234567, 32); |
| 209 |
} |
| 210 |
if (nc & 2) { |
| 211 |
unaligned_store_u16(c2, (uint16_t) _mm_extract_epi16(voutb2x01234567, 0)); |
| 212 |
c2 += 2; |
| 213 |
unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(voutb1x01234567, 0)); |
| 214 |
c1 += 2; |
| 215 |
unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(voutb0x01234567, 0)); |
| 216 |
c0 += 2; |
| 217 |
voutb2x01234567 = _mm_srli_epi32(voutb2x01234567, 16); |
| 218 |
voutb1x01234567 = _mm_srli_epi32(voutb1x01234567, 16); |
| 219 |
voutb0x01234567 = _mm_srli_epi32(voutb0x01234567, 16); |
| 220 |
} |
| 221 |
if (nc & 1) { |
| 222 |
*c2 = (int8_t) _mm_extract_epi8(voutb2x01234567, 0); |
| 223 |
*c1 = (int8_t) _mm_extract_epi8(voutb1x01234567, 0); |
| 224 |
*c0 = (int8_t) _mm_extract_epi8(voutb0x01234567, 0); |
| 225 |
} |
| 226 |
nc = 0; |
| 227 |
} |
| 228 |
} while (nc != 0); |
| 229 |
} |